In Content Aware this week: Sally's mystery family, AI can be quite stupid, and setting fire to $20bn
AI can't handle Sally and her 2 to 6 sisters
Positing a simple (for a human) written puzzle for 60 different Large Language Models to have a bash at easily showed the limitations of such systems: all of them got it wrong. The deviation in answers is a bit scary. If you want to see a fuller exchange between one of GPP's resident AI wranglers and ChatGPT 3.5, then click here. The word on AI Street is that GPT-4 gets it right first time, which is probably why Meta has it as a target.
Site rebuilds are generally accompanied by a feeling of optimism about how much better things will be once the project is completed. Yet, as we know, there's always a traffic penalty to such relaunches. Can it be avoided? No, says Google's John Mueller.
Eternal indemnity from Microsoft
Customers of Microsoft's AI assistant Copilot are to receive legal indemnity against any copyright claims for content generated using the system. MS must be very confident in its training data.
Corbidge comments on... the boring truth of being hands-on with AI systems
From "Ahh!" to "oh...", the arc of learning when using or experimenting with generative AI systems for real world publishing newsroom usage can be somewhat turbulent and disappointing. Our resident generative guy Rob reveals the pathway many of us have to take before realising that every AI has its limits - as well as their tricks and tweaks.
Do AI detection tools work?
No, according to OpenAI's recent documentation in responding to the question: "How can educators respond to students presenting AI-generated content as their own?" OpenAI should know, you'd assume, even admitting they've such a tool themselves.
Algorithm August: ripples and tremors
Google's most recent core search algorithm update - the one that can kill your traffic out of nowhere - is proving relatively hard to read, as they often are. Possibly for Google too. In the UK, and so likely elsewhere, marketplaces Amazon, Etsy, and Ebay have taken a hit, as have some large music and reference domains. The August Core update roll-out ended on the 7th September so as always check your data and see what jumps out as odd.
$20 billion you might as well have burned
Back to the dodgy ad game again: just in time for the Google megatrial, a new study for the US Association of National Advertisers is highly revealing. "There isn’t enough information in the hands of the people who are executing a programmatic campaign to make informed decisions about the right thing to do," leading to spectacular waste and misspend.
Someone else's test crash dummy
Google runs advertising experiments - this is known. Yet advertising experiments without the consent of the campaign's owners? Not sure anyone likes the sound of that.
Fair diffusion for all
Getty, Alamy, and other big names in the world of imagery have announced a joint Ethical AI Image Initiative called "The Fair Diffusion Program" that will see original creators properly compensated and credited for work used in AI training data. Can publishers co-operate similarly to text content?
How the NYT trained Googlebot to get more traffic
Great insight into how to work the Googlebot better.
Half a billion dollars on the way to local news projects.
Too many AI options? INMA helps
Giving some clarity in the newsroom.